Evaluasi Model Klasifikasi Motif Batik Lasem Menggunakan Xgboost, Lightgbm, Resnet50, dan Efficientnet-B0

Abstract

Batik Lasem is recognized for its rich visual complexity, making the identification of its motifs a challenging task that requires reliable image classification techniques. This study aims to evaluate the performance of four classification models LightGBM, XGBoost, ResNet50, and EfficientNetB0 in distinguishing five Batik Lasem motifs: Gunung Ringgit, Kricak/Watu Pecah, Latohan, Nyuk Pitu, and Seritan. The dataset consists of 5,879 images obtained from Mendeley Data and processed through several stages, including normalization, image resizing, and label structuring. For feature-based models, EfficientNetB0 was used as a feature extractor, whereas ResNet50 and EfficientNetB0 were implemented as end-to-end deep learning architectures. Model performance was assessed using 5-Fold Cross-Validation with accuracy, precision, recall, and F1-score as evaluation metrics.The experimental results indicate that LightGBM achieved the highest performance with an average accuracy of 0.9611, followed by XGBoost with an accuracy of 0.9556. Both boosting models demonstrated strong stability in handling the diverse texture characteristics of batik patterns. ResNet50 achieved an accuracy of 0.9029 and maintained solid capability in learning intricate visual features. In contrast, EfficientNetB0 obtained an accuracy of only 0.2343, suggesting that the model requires further optimization to effectively adapt to the high variability of Batik Lasem motifs. Overall, the findings reveal that boosting-based approaches are more effective for batik image classification, offering a solid foundation for future studies focused on deep learning optimization, enhanced augmentation strategies, or the exploration of more advanced architectures.